Improving the Robustness of Autonomous Vehicles Through Reinforcement Learning with Abnormally Behaving Traffic Actors

Most reinforcement learning (RL) methods for autonomous vehicles assume that other drivers behave rationally. However, in real-world traffic, drivers can act unpredictably due to distraction, fatigue, or aggression. This gap between training assumptions and real-world behavior can make autonomous systems less reliable. To improve robustness—the ability of an agent to perform well under unpredictable or difficult conditions—we propose a training approach that simulates abnormal driving behaviors. We introduce the Abnormal Wrapper, a mechanism that randomly assigns some vehicles in the simulation to perform erratic lane changes without regard to traffic context. This exposes the learning agent to unexpected situations during training. We train two Deep Q-Network (DQN) agents: one in a normal environment, and another in a modified environment with 50 % abnormal vehicles. Both are then tested under varying levels of traffic unpredictability, from 0 % to 75 % abnormal vehicles. Results show that the agent trained with abnormal traffic consistently performs better—it lasts longer on the road, earns higher rewards, and crashes less often. Notably, it also outperforms the normal agent even in ideal traffic, showing that exposure to noisy behaviors helps the agent learn safer, more adaptable driving strategies.